The Narrative Gap: Can LLMs Help Us Navigate Diverse Narratives Across Languages?
Across 14 LLMs, explicit requests for diverse perspectives still left substantial gaps in competing accounts.
The researchers introduced NarrativeBench, covering 1,333 news articles in 23 publication languages, 23 contested topics and 460 Reddit-grounded queries. They tested models across 24 query languages on six capabilities, including identifying missing narratives and preserving narrative coverage when questions seek confirmation. Access to information varied with both query framing and language. The authors warn that these gaps risk reinforcing echo chambers and linguistic filter bubbles.
HF Daily Papers' note
The researchers introduced NarrativeBench, covering 1,333 news articles in 23 publication languages, 23 contested topics and 460 Reddit-grounded queries. They tested models across 24 query languages on six capabilities, including identifying missing narratives and preserving narrative coverage when questions seek confirmation. Access to information varied with both query framing and language. The authors warn that these gaps risk reinforcing echo chambers and linguistic filter bubbles.
HF Daily Papers' note
score 4